Key facts at a glance

• Upscale, a Santa Clara, California-based networking startup backed by Nvidia, has introduced Token Fabric, an architecture that combines scale-up and scale-out networking for AI clusters.
• Token Fabric brings together Upscale's SkyFabriX scale-up silicon, scale-out systems built on Nvidia Spectrum-X Ethernet, and a common software layer called SkyOS with SkyCMD orchestration.
• General availability is planned for early 2027. Early-access and joint-validation programs are running now.
• Upscale emerged from stealth in September 2025 with a $100 million seed round and raised an additional $190 million in June. Nvidia joined that later round.
• The company has now raised about $500 million, employs more than 300 people, and is developing its SkyHammer custom scale-up switch chip. SkyFabriX is based on the SkyHammer architecture.
• Token Fabric extends existing protocols rather than replacing them. Scale-up traffic uses ESUN and standard Ethernet/IP, while scale-out traffic uses Ethernet with RoCE.
• Upscale says the architecture is aimed at neoclouds, enterprises, XPU vendors, and hyperscalers through three consumption models.

AI clusters rely on two distinct types of networking. Scale-up networking connects accelerators inside a rack, creating a tightly coupled domain for the fastest possible communication between GPUs or other processors. Scale-out networking connects racks across a data center, linking many smaller domains into a larger cluster. A single AI training or inference job can span both types of networking, but many organizations still design, deploy, and operate them as separate silos. That separation creates operational complexity, inconsistent telemetry, and limits on how efficiently a job can use available compute.

Networking startup Upscale is taking direct aim at that gap with Token Fabric. The architecture is designed to unify scale-up and scale-out networking under a common hardware and software approach. Instead of asking customers to stitch together separate fabrics, Token Fabric combines Upscale's own SkyFabriX scale-up silicon, scale-out systems built on Nvidia Spectrum-X Ethernet, and a shared software layer for operations. General availability is planned for early 2027, and early-access and joint-validation programs are already running.

The launch builds on a series of milestones for Upscale. The Santa Clara, California-based company emerged from stealth in September 2025 with a $100 million seed round. In June, it raised an additional $190 million and outlined plans for SkyHammer, its custom scale-up switch chip. SkyFabriX is based on the SkyHammer architecture. Nvidia, which supplies the Spectrum-X silicon used in the scale-out systems, joined the June round. Upscale has now raised about $500 million and has more than 300 employees.

Barun Kar, CEO of Upscale, said the need for a more integrated approach is growing as compute scales rapidly. In his view, networking has fallen behind compute, creating a bottleneck for AI clusters that need to move data among hundreds or thousands of accelerators. The problem is not just raw bandwidth. It is also latency, congestion management, protocol support, and the ability to operate a mixed environment without adding layers of complexity.

What is inside Token Fabric

Token Fabric is a combination of hardware and software for both scale-up and scale-out networking. The portfolio is built around several components that can be adopted together or separately.

• SkyFabriX silicon. SkyFabriX is the scale-up switch silicon. It supports ESUN, or Ethernet for Scale-Up Networking, as well as standard Ethernet/IP. It is designed for evolving protocols such as UALink, or Ultra Accelerator Link. Switch capacity is 115.2 Tbps, with a roadmap to multiple petabits per second.
• Scale-up switch trays. These trays come in custom and standard rack form factors. They pool GPUs and other accelerators into large scale-up domains so that accelerators inside a rack can communicate with low latency and high bandwidth.
• Scale-out systems. Upscale builds these around Nvidia Spectrum-X silicon at speeds from 400G and 800G up to 1.6T. With SkyOS, the systems integrate into Nvidia and other accelerator clusters.
• SkyOS and SkyCMD. SkyOS is the network operating system for both fabrics. It is based on SONiC, or Software for Open Networking in the Cloud, and optimized for AI protocols. SkyCMD is the orchestration and observability layer, providing a single orchestration plane across scale-up and scale-out.
• Consumption models. Customers can adopt silicon, systems, software, or the full stack. The full-stack option targets neoclouds and enterprises that do not have hyperscaler-size network teams. An integrated systems option targets XPU vendors, meaning makers of non-GPU accelerators, that want validated hardware without building a complete network stack. A third option provides silicon or custom systems with a software development kit for hyperscalers that run their own network operating system.

All three options draw on the same portfolio of silicon, systems, and software. Kar said the value comes from integration. In his description, the components are vertically stacked so that customers do not have to assemble a complete AI networking stack from unrelated parts.

How Token Fabric works

Token Fabric is built from the protocol layer up. Upscale is extending existing protocols rather than replacing them. Scale-up traffic uses ESUN and standard Ethernet/IP. Scale-out traffic uses standard Ethernet with RoCE, which carries remote direct memory access over Ethernet. The goal is to preserve compatibility with established networking practices while adding capabilities that AI workloads require.

Upscale is also contributing to the open projects underneath the stack. These include SONiC, extensions of the Switch Abstraction Interface for ESUN, the Ultra Ethernet Consortium, and UALink. That work matters because AI networking is still evolving. Customers want to avoid lock-in, but they also need hardware and software that can support new accelerator protocols as they mature.

Software ties the two fabrics together. SkyOS abstracts the underlying hardware and provides a control plane across the cluster. SkyCMD exposes that control plane through one layer. Multiple network elements sit behind it, and any compute platform in a heterogeneous cluster can use it to control the network. That means an operator does not need a separate management stack for every part of the fabric.

Kar said the operating system has to be lean, fast, flexible, and secure. The orchestration layer on top then enables capabilities such as heterogeneous compute, where different types of accelerators and compute platforms coexist in the same cluster. For AI operators, that heterogeneity is increasingly realistic. GPUs, custom XPUs, and other accelerators may all need to communicate through a shared network.

Why integration matters for AI networking

The traditional data center network was built around general-purpose servers and relatively predictable traffic patterns. AI clusters change the equation. Training jobs can create synchronized bursts of traffic across thousands of accelerators. Inference adds latency-sensitive request patterns and strict service-level expectations. Scale-up and scale-out networks have different design goals, but they serve the same job. Treating them as unrelated systems can leave performance on the table and increase the burden on operations teams.

Upscale's pitch is that integration can reduce that burden. A common software layer can provide one view of the network, one orchestration plane, and one set of policies across scale-up and scale-out. That is especially important for neoclouds and enterprises that want AI infrastructure without building a hyperscale network organization. It is also relevant for XPU vendors that need validated networking hardware but do not want to develop a full network stack.

For hyperscalers, the model is different. These customers often have large engineering teams and their own network operating systems. Upscale's third option gives them silicon or custom systems with a software development kit, allowing them to integrate the hardware into their own software environment. The common thread is that all three models draw from the same silicon, systems, and software portfolio.

Are tokens the new metric for network performance?

Network teams have traditionally tracked packets, throughput, loss, and jitter. The question for an AI network is whether tokens are the right unit of optimization. Kar argued that tokens are the right measure because users pay based on tokens today. He named time to first token, tokens per second, tokens per dollar, and tokens per watt as the measures that matter.

Optimizing those measures requires networking, especially in clusters that scale to hundreds of thousands of accelerators. A network that adds latency or drops traffic can slow token generation, reduce utilization, and raise cost per token. That is why Upscale frames Token Fabric around AI workload outcomes rather than only raw port speeds.

There is still a gap between traditional network operations and token-focused optimization. According to Kar, the main components of that gap are latency, bandwidth, and scale-up features in the silicon. On the software side, predictive analytics and telemetry keep machines running. The network has to detect problems early and route around them before they become job-level failures.

The other side of the challenge is optimizing the network with agentic AI. Upscale has built agent-based operations into the platform. Kar said an agent is always looking at the network, trying to figure out whether cables will fail or optics will fail or whether there is congestion, and the agents are looking for ways around it. That continuous monitoring and remediation model reflects a broader shift in AI infrastructure: the network is not just a static pipe, but an active part of how efficiently tokens are produced.

Source:Network World News